Bio: Machine Learning Research Engineer. Gimme the data.
Location: London, UK
🏄♂️ OscarSavolainenDR
Machine Learning Research Engineer
My 3 great engineering loves are analyzing data, neural network quantization, and unit testing of complex use cases. At the risk of sounding like a complete nerd, I am super into working on very complex problems that have a strong coding aspect. It's how I get into flow state. At the moment, that is building a neural network quantization library with never-before-seen tools and techniques, with an ultra-fast Rust backend. And yes, it interfaces seamlessly with PyTorch.
🧰 Languages and Tools
📺 Latest YouTube Videos
📊 Stats
👨💻 Oscar's Coding Journey
I first got started in coding as a means of enabling me to do what I love: data analysis. I go cuckoo for data, and coding was a way to enable gathering, transforming, and visualizing numbers. Over time I ended up using more and more advanced techniques. When I was doing my PhD in neurotechnology at Imperial College, to tackle complex biological data, I had to start getting the big algorithms involved: Machine Learning. Before my PhD even ended, I started working professionally as an ML Researcher, and grew to love ML for itself: understanding how it learns transforms, the subtleties of forward and backwards passes, and most of all, how it reacts when we throw a sackful of wrenches into the motor of the algorithm when we do quantization. At the moment, I'm excited to be educating others on neural network quantization and building my own quantization library, while continuing my journey of diving down into computational optimization, low-level languages such as Rust, and playing with various LLM use cases.
A minimalist extension to PyTorch's quantization library, which improves QAT training speed, adds custom visual quantization debugging tools, and robust unit testing. Natively compatible with both Eager and Graph mode quantization.
A Monte-Carlo based statistical significance test for inter-frequency power correlations in non-stationary time-series. Accounts for intra-frequency autocorrelation, inter-frequency non-dyadicity, and controls the FDR for multiple testing under dependency.
Extremely minimal Python code on how to query a Huggingface (embedding) model running on an infinity server on a Runpod instance. For educational purposes.
A website I made for real-estate analysis. It webscrapes publicly availably to-rent listings and calculates their profitability as STLs based on surrounding comparables. It summarizes all of the results per listing that one can filter through, and provides a downloadable `.csv` file per listing with the full analysis per listing.